RPEU-Net: Residual Parallel Enhanced Attention Network for Left Atrial Segmentation

Accurate left atrial segmentation is essential for improving cardiovascular diagnosis and therapeutic strategies. The adoption of fully convolutional networks, particularly those based on the U-shaped network (U-Net) architecture, has substantially improved semantic segmentation accuracy, becoming predominant in medical segmentation tasks. Nevertheless, the left atrium, characterized by its complex and non-rigid motion, presents substantial challenges for existing segmentation models, which often fail to fully utilize channel feature information and suffer from redundant semantic information extraction, thereby compromising segmentation accuracy. To address these limitations, the RPEU-Net network model for left atrial segmentation is proposed. Initially, the residual parallel convolution unit (RPCU) is initially incorporated to improve the model’s capability in capturing and representing complex feature information. Subsequently, the proposed hybrid enhanced attention module (HEAM) mitigates the shortcomings of traditional attention mechanisms, significantly improving the screening capability of relevant feature information and demonstrating robust adaptability to noise and redundant information. Finally, the network employs a lightweight architectural design to reduce redundancy while maintaining computational efficiency. Evaluation experiments conducted on two left atrial datasets indicate that RPEU-Net achieves superior results, exhibiting higher segmentation performance and potential compared to existing methods.